改进YOLO的车辆检测算法  

Improved YOLO vehicle detection algorithm

作　　者：刘肯 何姣姣 张永平[2] 姚拓中[2] 常志国[1] LIU Ken;HE Jiaojiao;ZHANG Yongping;YAO Tuozhong;CHANG Zhiguo(School of Information Engineering,Chang’an University,Xi’an 710064,China;College of Telecom,Ningbo University of Technology,Ningbo 315211,China)

机构地区：[1]长安大学信息工程学院,陕西西安710064 [2]宁波工程学院电信学院,浙江宁波315211

出　　处：《现代电子技术》2019年第13期47-50,共4页Modern Electronics Technique

基　　金：国家自然科学基金(61771270);浙江省自然科学基金(2017A610109);浙江省自然科学基金(LQ15F020004)~~

摘　　要：YOLO目标检测算法在进行目标检测和识别时具有识别精度高、检测速度快的特点。但也存在明显的问题,由于网络采用的结构是端对端模型,没有经过预选框进行匹配预选,直接进行回归产生最后的结果框,所以存在定位不准的问题。同时,YOLO网络将图片整体resize到固定的尺寸后,进行网格划分,当单个网格存在不止单个目标时,容易出现漏检的情况。DenseNet网络使用一种全新的网络结构,结合前面特征层的信息,在一定程度上提升了对物体的检测精度。在此基础上提出YOLO-D算法,结合前面特征层的信息,在不影响检测速度的同时,提高车辆检测精度,与此同时使得定位有所改善。The YOLO target detection algorithm has the characteristics of high recognition accuracy and fast detection speed for target detection and recognition,but also has obvious problems. The network adopts an end-to-end model,and directly generates the final result framework after regression without matching and pre - selection of pre - selection framework,so the positioning is inaccurate. The YOLO network can resize the entire image to a fixed size,and divide the mesh. When there is more than one target in a single mesh,it is likely to result in the missed detection. A brand-new network structure is used in DenseNet network,and combined with the information of the front feature layer to improve the detection accuracy of the object to a certain extent. Thus,the proposed YOLO-D algorithm combined with the information of the front feature layer can improve the vehicle detection accuracy and positioning accuracy while maintaining the vehicle detection speed.

关 键 词：YOLO 端对端模型 DenseNet 车辆检测 YOLO-D 检测精度 

分 类 号：TN911.73-34[电子电信—通信与信息系统]